add function to combine all tasks metadata into tasks.json

[zeke]

This PR updates the inference-codegen script to build a JSON file containing all the metadata of all the tasks.

The goal is to make this data more portable, so it can be easily consumed in different contexts outside of the huggingface.js codebase, e.g. for Replicate model classification.

To test:

$ pnpm run inference-codegen

Here's an example of the current output: https://gist.github.com/zeke/460a58b7aa50e305072415844a335209

To Do

• Update inference-codegen script to output all task data as tasks.json
• Dereference the JSON schema so everything is included.
• Make sure summary is included. e.g. "Keypoint detection is the task of identifying meaningful distinctive points or features in an image."
• git-ignore generated tasks.json file.
• ???

[zeke]

@SBrandeis @julien-c @Wauplin 👋🏼

I finally got around to bumping this forward a bit. My goal is to use this to classify all the models on Replicate.

Could use a little help with the dereferencing the JSON $refs.. I'm not familiar with quicktype and my initial attempts failed.

[Wauplin]

Hey @zeke , thanks for looking into this! Just to be sure, the goal of this PR is to generate a single tasks.json with all definitions from all input/output/output_stream from all tasks defined in https://github.com/huggingface/huggingface.js/tree/main/packages/tasks/src/tasks. Am I correct? No extra information added to it? (just want to be sure about the goal of it)

[zeke]

Hey @zeke , thanks for looking into this! Just to be sure, the goal of this PR is to generate a single tasks.json with all definitions from all input/output/output_stream from all tasks defined in https://github.com/huggingface/huggingface.js/tree/main/packages/tasks/src/tasks. Am I correct? No extra information added to it? (just want to be sure about the goal of it)

Yep! That's the goal!

[zeke]

Poking around the codebase, I see other code paths where more task types are mentioned. Here's one:

huggingface.js/packages/tasks/src/tasks/index.ts

Line 132 in 1aa1c3f

"image-to-video": ["diffusers"],

For example, image-to-video is not present in the code I'm generating here, but it is present in that file ☝🏼

Let me know if there's a better way to structure this to be using the most complete and up-to-date set of Tasks.

[zeke]

Ideally I also want the generated file to include these prose summary strings for each Task:

huggingface.js/packages/tasks/src/tasks/image-to-image/data.ts

Line 103 in 1aa1c3f

"Image-to-image is the task of transforming an input image through a variety of possible manipulations and enhancements, such as super-resolution, image inpainting, colorization, and more.",